AiX 環球 AI 情報 Global AI Intelligence

第 16 期 2026 年 10 月 13 日（星期二） 07:30 出刊 10 條情報 · 22 原文來源連結 RSS 過往期數

今日十條 · 2026 年 10 月 13 日同一天，代理被要求付費、被要求問責、也被要求供電：能力競賽的三條新底線

今日主編判斷

十月十二日這一天，環球人工智能產業同時在三條底線上收緊：前沿模型首度以訂閱制直接向個人收費，代理的安全問責由自願承諾轉為法定時限，而支撐代理的算力則由電力與融資兩端設下硬上限。我們認為，企業決策者過去一年關注的「模型有多強」正在讓位於三個更實際的問題：這一層能力要付多少錢、出事時由誰在多少小時內上報、以及支撐它的電與債能否撑得住。誰先把這三條底線寫進採購與治理流程，誰就能在下一個預算週期少付一次學費。On 12 October, the global AI industry tightened three baselines at once: a frontier model charged individuals directly by subscription for the first time, agent safety accountability moved from voluntary pledges to statutory clocks, and the compute behind those agents met hard ceilings on both power and financing. Our view is that the question which preoccupied decision-makers over the past year — how capable is the model — is giving way to three far more practical ones: what this layer of capability costs, who must report an incident and within how many hours, and whether the electricity and debt behind it can hold. Those who write these three baselines into procurement and governance first will pay one less tuition fee in the next budget cycle.

今日十條

本欄由 AIX Society 編輯部整理公開資料後撰稿，每條均附原始來源連結；事實與觀點分列，觀點屬本會判斷。

按受眾篩選

Meta 發布 Muse Wonder，首次以每月 20 至 100 美元訂閱直接向個人收費Meta Releases Muse Wonder, Charging Individuals Directly via US$20–100 Monthly Subscriptions for the First Time

影響力✓ 已核實Verified已在發生Happening now企業決策者Business owners資訊科技IT

事件: Meta 於十月十二日發布 Muse Wonder 與 Muse Wonder Full，由 Alexandr Wang 領導的 Meta Superintelligence Labs 開發。Muse 應用程式用戶須付每月 20 或 100 美元訂閱方可使用；最強版本 Muse Wonder Full 僅限 Muse Code 訂閱者，月費 50 或 200 美元。API 定價為每百萬輸入代幣 4 美元、輸出 20 美元。Meta released Muse Wonder and Muse Wonder Full on 12 October, developed by Meta Superintelligence Labs under Alexandr Wang. Muse app users must pay a US$20 or US$100 monthly subscription to access it; the strongest variant, Muse Wonder Full, is limited to Muse Code subscribers at US$50 or US$200 a month. API pricing is US$4 per million input tokens and US$20 per million output tokens.
背後: Meta 的營收歷來幾乎全部來自廣告，出售模型能力是它第一次在廣告之外建立直接收入。把最強版本鎖在付費牆後、只把較小模型放上 API，說明它賭的是消費者付費意願，而非開發者分發。Meta's revenue has historically come almost entirely from advertising, so selling model capability is its first attempt to build direct revenue outside ads. Locking the strongest variant behind a paywall while putting only the smaller model on the API shows it is betting on consumer willingness to pay rather than developer distribution.

顧問觀點

我們認為，前沿模型由「按用量計」轉向「按人頭訂閱」，會令企業的 AI 預算結構由變動成本變成固定成本，談判籌碼亦從 API 折扣轉向席位數與續約週期。採購部門應及早把訂閱制納入年度預算模型。Our view is that the shift from usage-based pricing to per-seat subscriptions will turn part of the enterprise AI budget from a variable cost into a fixed one, and move negotiating leverage from API discounts towards seat counts and renewal cycles. Procurement teams should fold subscription models into their annual budgeting early.

對日常工作的影響: 對財務與採購而言，AI 開支由可隨時調整的用量帳單，變成須按年鎖定的席位合約；IT 則要多一層席位分配與離職回收的管理工序。For finance and procurement, AI spending moves from an adjustable usage bill to a seat contract locked in for the year; IT gains an extra layer of seat allocation and offboarding reclamation work.
如何改善: 本週請財務把你公司現有的 AI 工具按「用量計費」與「席位訂閱」分兩欄列出，標出各自年度總額，看清固定成本佔比。This week, ask finance to list your existing AI tools in two columns — usage-billed and seat-subscribed — with the annual total for each, so you can see the share that is fixed cost.

來源CNBC（2026-10-12）↗AI Weekly（2026-10-12）↗TPS Report（2026-10-12）↗

英國公布 AI 風險方案：研究把代理失控與事故上報寫入網絡安全法案UK Unveils AI Risk Package: Examining Loss-of-Control and Incident Reporting Under the Cyber Security Bill

影響力✓ 已核實Verified已在發生Happening now法務合規Legal & compliance資訊科技IT企業決策者Business owners

事件: 英國政府十月十二日公布一套 AI 風險措施，由內閣辦公室、AI 安全研究所、國家網絡安全中心及數碼文化媒體體育部聯合發布。當局稱前沿 AI 網絡防禦先導計劃已協助發現並修復逾一千個政府系統潛在漏洞，正研究把嚴重 AI 事故與「近失誤」的識別與上報納入《網絡安全與韌性法案》。The UK government announced a package of AI risk measures on 12 October, issued jointly by the Cabinet Office, the AI Security Institute, the National Cyber Security Centre and the Department for Digital, Culture, Media and Sport. It said a frontier-AI cyber defence pilot has helped find and fix more than 1,000 potential vulnerabilities in government systems, and that it is examining how the identification and reporting of serious AI incidents and near misses could be brought into the Cyber Security and Resilience Bill.
背後: 該先導計劃以六階段代理鏈掃描政府公開程式碼，另有一組先跑傳統工具再疊加三個模型階段，全月共四百零七項發現，代幣成本僅一萬三千英鎊。以極低成本換取可審計的安全收益，正是當局願意把上報責任制度化的底氣。The pilot ran public government code through a six-stage agent chain, while another team ran conventional scanners first and layered three model stages on top, producing 407 findings in a month at a token cost of only £13,000. Low cost against auditable security gains is precisely what gives the government confidence to institutionalise reporting duties.

顧問觀點

我們認為，這份方案最值得注意的不是罰則，而是「近失誤」也要上報這一設計。它把安全責任由結果問責前移到過程留痕，意味企業日後要證明的不只是沒出事，而是每次異常都有記錄、有覆核。Our view is that the most notable feature is not the penalties but the inclusion of near misses in reporting. It moves safety accountability from outcome to process traceability, meaning firms will later have to prove not only that nothing went wrong, but that every anomaly was logged and reviewed.

對日常工作的影響: 對法務與 IT 而言，日後須為代理異常建立可重建的日誌與內部上報流程；前線運維則要習慣把「未釀成事故的異常」當作正式紀錄處理。For legal and IT, agent anomalies will need reconstructable logs and an internal reporting path; frontline operations will have to treat anomalies that did not cause harm as formal records.
如何改善: 本週請 IT 選定一條已在運行的代理流程，試填一份「近失誤記錄表」：時間、觸發、影響、誰覆核，先跑一次看流程是否行得通。This week, ask IT to pick one live agent workflow and trial a near-miss log — time, trigger, impact, who reviewed it — to see whether the process holds up once.

來源Unite.AI（2026-10-12）↗CTOL（2026-10-12）↗GOV.UK（2026-10-12）↗

微軟發布 Maia 300 自研推理晶片，稱每兆瓦代幣量較主流 GPU 高四成Microsoft Unveils Maia 300 In-House Inference Chip, Claiming 40% More Tokens per Megawatt Than Leading GPUs

影響力✓ 已核實Verified6–12 個月6–12 months資訊科技IT企業決策者Business owners

事件: 微軟於十月十二日在聖何塞 OCP 全球峰會發布第三代自研 AI 加速器 Azure Maia 300，將於二○二七年開始量產部署，並計劃擴展至數吉瓦級。公司稱該晶片每兆瓦可處理代幣量較領先 GPU 系統高約四成，單一機櫃可容 96 顆加速器、功耗約 167 千瓦，代幣吞吐量最高達 Maia 200 的七倍。Microsoft unveiled its third-generation in-house AI accelerator, Azure Maia 300, at the OCP Global Summit in San Jose on 12 October. Production deployment begins in 2027 with plans to scale to multi-gigawatt capacity. The company says it delivers about 40% more tokens per megawatt than leading GPU-based systems; a single rack holds 96 accelerators drawing roughly 167 kilowatts, with up to seven times the token throughput of Maia 200.
背後: 微軟長期依賴英偉達與 AMD 供應訓練晶片，但推理是用電與成本最持續的一環。把自研晶片定位為「主要推理系統」並只用於自家服務、不對外出租，等於先在內部把算力成本壓低，再決定是否讓利給客戶。Microsoft has long depended on Nvidia and AMD chips for training, but inference is the most persistent cost and power draw. Positioning its own silicon as the primary inference system and reserving it for internal use rather than renting it out lets Microsoft compress compute costs internally first, then decide whether to pass savings to customers.

顧問觀點

我們認為，這則消息的重點不在四成這個數字，而在「主要推理系統」六個字。它代表超大規模雲商開始以自研晶片重訂推理成本線，企業日後比較雲服務時，要問的不再只是每 token 報價，而是對方用什麼晶片供電。Our view is that the key point is not the 40% figure but the phrase 'primary inference system'. It marks hyperscalers beginning to redraw the inference cost line with their own silicon; when comparing cloud services, enterprises will need to ask not only the per-token price but which chip the provider is powering it with.

對日常工作的影響: 對 IT 與財務而言，雲服務的成本曲線將愈來愈取決於供應商的晶片自給率；短期採購談判應把長約的價格調整條款與供應商硬體路線掛鈎。For IT and finance, the cloud cost curve will increasingly depend on a provider's chip self-sufficiency; near-term procurement talks should tie long-contract price adjustment clauses to the vendor's hardware roadmap.
如何改善: 本週請 IT 向主要雲供應商書面查詢：現行方案所用推理硬體為何、未來十二個月會否更多轉向自研晶片，並把答案存檔作議價依據。This week, ask IT to put a written question to your main cloud vendors: which inference hardware the current plan runs on, and whether more move to in-house silicon over the next twelve months — and file the answers as negotiating evidence.

來源AI in Context（2026-10-12）↗Trading Dashboard（2026-10-12）↗

階躍星辰 Step 5 Preview 接入騰訊 WorkBuddy，獨家限免兩週StepFun's Step 5 Preview Lands Inside Tencent WorkBuddy with an Exclusive Two-Week Free Window

影響力✓ 已核實Verified已在發生Happening now資訊科技IT前線員工Frontline staff

事件: 阶跃星辰於十月十二日宣布，旗艦模型 Step 5 Preview 正式接入騰訊全場景 AI 辦公工作台 WorkBuddy，由十月十二日至二十五日在該平台獨家限免。該模型採稀疏 MoE 架構，總參數 6000 億、每代幣激活約 270 億，支援一百萬代幣上下文，原生支援文本與視覺輸入，並計劃於十月十五日開源完整權重。StepFun announced on 12 October that its flagship Step 5 Preview model is now available inside Tencent's all-scenario AI office workspace, WorkBuddy, with exclusive free access on the platform from 12 to 25 October. The model uses a sparse mixture-of-experts architecture with 600 billion total parameters and about 27 billion activated per token, supports a one-million-token context window, accepts text and visual input natively, and is scheduled to have its full weights open-sourced on 15 October.
背後: 內地辦公代理賽道競爭加劇，模型商需要真實辦公入口來驗證能力，辦公平台則需要更多模型選擇來綁住用戶。限免兩週是典型的雙向試用期：模型商換取真實任務數據，平台換取用戶停留，期滿後是否留存取決於實際表現。Competition in China's office-agent segment is intensifying: model makers need real office entry points to validate capability, while office platforms need more model choices to retain users. A two-week free window is a classic two-way trial — the model maker gains real task data, the platform gains user stickiness, and retention after it closes depends on actual performance.

顧問觀點

我們認為，這類「模型進駐辦公平台」的合作會令企業用戶的選擇權表面上增加、實際上轉移到平台手裡。真正要問的是：當平台可以隨時更換底層模型，企業的輸出品質、可重現性與資料邊界由誰保證。Our view is that such 'model-into-platform' deals appear to widen enterprise choice while actually shifting it to the platform. The real question is this: when a platform can swap the underlying model at will, who guarantees output quality, reproducibility and data boundaries for the enterprise?

對日常工作的影響: 對 IT 與營運而言，同一份工作可能因底層模型更換而結果不同；日常工序需要建立版本記錄，避免交付物在事後無法重現。For IT and operations, the same task may produce different results when the underlying model changes; daily workflows need version records so deliverables remain reproducible after the fact.
如何改善: 本週請選一條常用代理工序，連續三天記錄所用模型版本與輸出，建立一份最小可行的「模型版本—輸出」對照紀錄。This week, pick one routinely used agent workflow and for three days record the model version and output, building a minimal model-version-to-output record.

來源Pandaily（2026-10-12）↗騰訊新聞（2026-10-12）↗

博通為 OpenAI 定製晶片籌組逾五百億美元融資，SpaceX 另籌四百億購英偉達晶片Broadcom Arranges Over US$50bn to Back OpenAI's Custom Chips as SpaceX Seeks US$40bn for Nvidia Hardware

影響力✓ 已核實Verified6–12 個月6–12 months企業決策者Business owners資訊科技IT

事件: 據《華爾街日報》報道，博通正為其與 OpenAI 聯合開發的定製 AI 晶片籌組超過五百億美元融資，Apollo 與 Blackstone 已就參與洽談；SpaceX 則商談約四百億美元融資以採購英偉達處理器，當中一百億為銀行貸款、三百億為投資級債券，料二○二七年完成。甲骨文亦與 Apollo 及高盛磋商同類安排。According to the Wall Street Journal, Broadcom is arranging more than US$50 billion in financing for the custom AI chip it is developing with OpenAI, with Apollo and Blackstone in talks to participate; SpaceX is discussing roughly US$40 billion, split between US$10 billion in bank loans and US$30 billion in investment-grade bonds, to buy Nvidia processors, expected to close in 2027. Oracle is also in talks with Apollo and Goldman Sachs on a similar arrangement.
背後: 傳統銀行受資本充足率約束，對折舊快、收入依賴新興市場的資產態度保守，於是私募信貸填補空隙。當 Apollo 同時出現在幾宗 AI 融資中，說明替代資產管理機構已把 AI 算力視為可融資的資產類別。Traditional banks, constrained by capital adequacy rules, stay cautious about assets that depreciate fast and rely on a nascent market for revenue, so private credit fills the gap. Apollo appearing in several AI financings at once shows alternative asset managers now treat AI compute as a financeable asset class.

顧問觀點

我們認為，這一波融資真正改變的是風險落點：還款綁定未來的雲收入，而非現有資產。企業在簽署與算力供應商的長約前，宜先確認對方的融資結構，因為一旦行業下行，這類貸款的壓力會先反映在服務條款上。Our view is that what this wave really changes is where the risk sits: repayment is tied to future cloud revenue rather than existing assets. Before signing long contracts with compute suppliers, enterprises should first understand their counterparty's financing structure, because in a downturn the pressure from such debt shows up first in service terms.

對日常工作的影響: 對財務與採購而言，算力供應商的信用狀況將直接影響服務穩定性；長約應加入供應商融資條款變動的通知與退出安排。For finance and procurement, a compute supplier's credit standing now bears directly on service stability; long contracts should include notification and exit terms when the supplier's financing arrangements change.
如何改善: 本週請財務在與算力或雲服務供應商簽約前，加查一項：該供應商是否有以其 AI 相關資產作抵押的大額融資，並記錄到期年份。This week, before signing with a compute or cloud provider, ask finance to check one extra item: whether the supplier carries large financing secured against its AI-related assets, and note the maturity years.

來源The Data Center Digest（2026-10-12）↗Daily 8 News（2026-10-12）↗Flux Aeriox（2026-10-12）↗

新加坡資本投資韓國浦項 40MW AI 數據中心，規劃擴至 300MWSingapore Capital Invests in a 40MW AI Data Centre in Pohang, Korea, with Plans to Scale to 300MW

影響力✓ 已核實Verified6–12 個月6–12 months資訊科技IT企業決策者Business owners

事件: 新加坡私募機構 SC Capital Partners 於十月十二日宣布，與韓國 AI 基建業者 Neo AI Cloud 合作，完成浦項一座 40MW AI 數據中心的開發投資，預計二○二七年底竣工，後續規劃擴建 260MW，園區總規模可望達 300MW。項目已取得電力供應核准與建築許可並動工，液冷設施可支援英偉達下一代 Vera Rubin GPU。Singapore private equity firm SC Capital Partners announced on 12 October that, with Korean AI infrastructure operator Neo AI Cloud, it has completed an investment in the development of a 40MW AI data centre in Pohang. Completion is expected by end-2027, with a further 260MW planned, bringing the campus to a potential 300MW. The project has secured power approval and a construction permit, has broken ground, and its liquid-cooling infrastructure is ready for Nvidia's next-generation Vera Rubin GPUs.
背後: 亞太不動產資本正把資金由傳統物業轉向高功率 AI 設施，而電力取得與租戶預租是這類項目最關鍵的兩個變數。該項目同時鎖定兩者，說明開發商知道融資方要看的不是概念，是可交付的電與已簽的租約。Asia-Pacific real estate capital is shifting from conventional property into high-density AI facilities, where power access and pre-leasing are the two decisive variables. This project locks down both, showing developers know financiers want deliverable electricity and signed leases rather than a concept.

顧問觀點

我們認為，這宗交易揭示 AI 基建的競爭已從「誰有晶片」轉向「誰拿到電與地」。對有意自建算力的企業而言，選址階段的電力談判能力，比伺服器規格更能決定項目的成敗與時程。Our view is that this deal reveals AI infrastructure competition has moved from who has the chips to who secures the power and land. For enterprises considering building their own compute, the ability to negotiate power at the site-selection stage will determine project success and timing more than server specifications.

對日常工作的影響: 對負責選址與基建的 IT 與營運主管而言，電力審批與租戶承諾的進度表，將直接決定自建或租用算力的決策時點。For IT and operations leads handling siting and infrastructure, the timetable for power approval and tenant commitments will directly set the decision point for building or renting compute.
如何改善: 本週請負責基建的同事把「電力審批」與「租戶／容量承諾」列為自建算力評估表的兩個一級門檻，並為每項訂下最早可確認日期。This week, ask whoever handles infrastructure to make power approval and tenant/capacity commitments the two first-tier gates in your build-versus-rent assessment, each with an earliest confirmable date.

來源優分析（2026-10-12）↗騰訊新聞（2026-10-12）↗

南韓啟動 800 兆韓圜西南部半導體與 AI 群集，造船業同步投入四千億韓圜Korea Launches an 800 Trillion Won Southwestern Semiconductor and AI Cluster as Shipbuilders Pledge 400 Billion Won

影響力✓ 已核實Verified3 年3 years企業決策者Business owners人力資源HR

事件: 南韓產業通商資源部長金正官於十月十二日視察光州與木浦，推進以可再生能源支撐的西南部半導體與 AI 群集，背景是規模達八百兆韓圜（約五千九百八十九億美元）的投資計劃。同日造船 M.AX 聯盟會議敲定，HD 現代三湖與大亞產業合共承諾投入四千億韓圜，在船廠建立 AI 燈塔工廠。South Korea's Minister of Trade, Industry and Energy, Kim Jung-kwan, visited Gwangju and Mokpo on 12 October to advance a renewables-powered semiconductor and AI cluster in the southwest, against the backdrop of an 800 trillion won (about US$598.9 billion) investment plan. At a shipbuilding M.AX Alliance meeting the same day, HD Hyundai Samho and Daea Industries jointly pledged 400 billion won to build AI lighthouse factories at shipyards.
背後: 這是南韓「五極三特」增長引擎計劃的落地環節。政府先解決土地與電力等硬基建，再把大型船廠的 AI 轉型經驗擴散至中小型供應商，試圖以區域群集方式把 AI 紅利留在本土製造業。This is the implementation phase of Korea's Five-Pole, Three-Special growth engine plan. The government first resolves hard infrastructure such as land and power, then extends large shipyards' AI transformation experience to smaller suppliers, seeking to keep AI gains within domestic manufacturing through regional clustering.

顧問觀點

我們認為，這類「政府先鋪電與地、企業再投入」的模式，值得本地製造業參考。當 AI 轉型由單一企業的資本支出變成區域公共基建，中小企的導入門檻才會真正下降。Our view is that this model — government laying power and land first, then enterprises investing — is worth studying for local manufacturers. Only when AI transformation shifts from a single firm's capital expenditure to regional public infrastructure does the entry barrier for SMEs genuinely fall.

對日常工作的影響: 對製造業營運與 HR 主管而言，區域 AI 基建成熟後，自動化設備與技能培訓的取得成本會同步下降，人力規劃需提前考慮工序重組。For manufacturing operations and HR leads, as regional AI infrastructure matures, the cost of acquiring automation equipment and skills training falls in tandem, and workforce planning must bring process restructuring forward.
如何改善: 本週請營運主管列出你工廠內重複度最高的三道工序，評估若同類 AI 基建在本地落成，哪一道最先適合導入自動化。This week, ask your operations lead to list the three most repetitive processes in the plant and assess which would be the first candidate for automation if similar AI infrastructure were built locally.

來源BigGo Finance（2026-10-12）↗AJU PRESS（2026-10-12）↗

英國資訊專員就代理式 AI 展開證據徵集，並確認十家模型商已作資料保護調整UK ICO Opens Agentic AI Call for Evidence and Confirms Ten Model Developers Have Made Data Protection Changes

影響力✓ 已核實Verified已在發生Happening now法務合規Legal & compliance資訊科技IT

事件: 英國資訊專員公署十月八日發布報告，指亞馬遜、Anthropic、蘋果、Cohere、DeepSeek、Google、Meta、微軟、OpenAI 與 Stability AI 十家基礎模型開發商，已作出或承諾作出資料保護調整，涵蓋透明度資訊、權利行使機制與保障評估；同日就代理式 AI 展開為期六週的證據徵集，至十一月二十日截止。The UK Information Commissioner's Office published a report on 8 October stating that ten foundation model developers — Amazon, Anthropic, Apple, Cohere, DeepSeek, Google, Meta, Microsoft, OpenAI and Stability AI — have made or committed to data protection changes covering transparency information, rights-exercise mechanisms and safeguard assessments. It opened a six-week call for evidence on agentic AI the same day, closing on 20 November.
背後: 監管機構今次走的是資料保護法而非 AI 專法路線，因此拿到的是經談判取得的約束性承諾，而非罰款。這種做法的好處是可即時生效，代價是範圍只覆蓋與個人資料相關的部分。The regulator has taken the data protection route rather than AI-specific legislation, which is why it obtained negotiated, binding commitments instead of fines. The advantage is immediate effect; the cost is that the scope covers only what touches personal data.

顧問觀點

我們認為，「代理有自主性不是合規不足的藉口」這句話，是這一輪監管最值得企業記住的一句。它意味自主程度不能再作為責任豁免的理由，企業與供應商的合約須重新劃分代理行為的責任歸屬。Our view is that the regulator's line — that an agent's autonomy is no excuse for poor compliance — is the single sentence enterprises should remember from this round. It means autonomy can no longer be used to disclaim responsibility, and contracts with suppliers must reallocate liability for agent actions.

對日常工作的影響: 對法務與負責供應商管理的採購而言，現有 AI 合約須補上代理行為的責任條款與資料處理範圍，否則風險會默認留在採購方。For legal and procurement managing vendors, existing AI contracts must add liability clauses for agent actions and data processing scope, or the risk defaults to the buyer.
如何改善: 本週請法務挑出三份最主要的 AI 供應商合約，逐一標示「代理自主行為的責任歸屬」是否已有明確條款，沒有的列入待補清單。This week, ask legal to pull the three most important AI supplier contracts and mark whether each clearly assigns liability for autonomous agent actions, listing the gaps for amendment.

來源Beinsure（2026-10-12）↗Licentium（2026-10-12）↗

招聘數據顯示入門門檻收窄：澳洲初階職位僅 4.9% 涉及 AI，主管級達 31.1%Hiring Data Shows a Narrowing Entry Gate: Only 4.9% of Australian Entry-Level Roles Involve AI, Against 31.1% at Lead Level

影響力✓ 已核實Verified已在發生Happening now人力資源HR企業決策者Business owners

事件: Verellax 指數分析二○二六年第三季七個市場、二百八十四家僱主共五萬七千一百三十四個職位。澳洲初階職位中僅 4.9%（八十九個）涉及 AI，主管級為 31.1%（六百一十六個），相差逾六倍；全部市場初階職位涉及 AI 的比例為 15.6%。另據 Indeed 招聘實驗室數據，美國軟件開發職位自二○二五年二月以來上升近 15%，但初級崗位減少。The Verellax Index analysed 57,134 roles advertised by 284 employers across seven markets in Q3 2026. Only 4.9% of Australian entry-level roles (89 of 1,812) involve AI, against 31.1% at lead level (616 of 1,981) — a gap of more than six times; across all markets, 15.6% of entry-level roles involve AI. Separately, Indeed Hiring Lab data shows US software development postings up nearly 15% since February 2025, while junior roles have declined.
背後: 僱主一邊說 AI 會改變每個人的第一份工，一邊在職位描述裡幾乎不寫 AI。這種落差說明目前的 AI 職能仍集中在有經驗者身上，因為部署與覆核代理需要判斷力，而判斷力正是初階職位最缺的。Employers say AI will change everyone's first job while barely writing AI into those job ads. The gap shows AI duties currently sit with experienced staff, because deploying and reviewing agents requires judgement — precisely what entry-level roles lack.

顧問觀點

我們認為，AI 對招聘的影響目前不是「取代職位」，而是「重新分配入門機會」。企業若繼續只在資深職位加註 AI 要求，三年後會發現沒有內部人才承接代理治理的工作。Our view is that AI's impact on hiring is currently not replacing roles but reallocating entry opportunities. If firms keep adding AI requirements only to senior postings, in three years they will find no internal pipeline to take on agent governance.

對日常工作的影響: 對 HR 與主管而言，招聘條件需明確分出「須具備」與「可在職培訓」的 AI 技能，並為初階職位設計可累積的代理監督經驗。For HR and line managers, job requirements must separate AI skills that are essential from those trainable on the job, with entry roles designed to accumulate agent-supervision experience.
如何改善: 本週請 HR 抽樣審視最近五份招聘廣告，把當中所有「AI 相關要求」逐條標為必要或可培訓，重新調整初階職位的入職門檻。This week, ask HR to sample the last five job ads, mark every AI-related requirement as essential or trainable, and recalibrate the entry threshold for junior roles.

來源The National Tribune（2026-10-13）↗WWC One Media（2026-10-12）↗AI Market Pulse（2026-10-12）↗

微軟總裁史密斯：AI 受監管程度不及牛奶與洗衣機，倡設緊急制動與責任條款Microsoft President Smith: AI Is Less Regulated Than Milk or Washing Machines, Urging Kill Switches and Liability Duties

影響力✓ 已核實Verified已在發生Happening now資訊科技IT法務合規Legal & compliance企業決策者Business owners

事件: 微軟總裁布拉德·史密斯於十月十二日接受 BBC 訪問時表示，人工智能受監管程度甚至不如牛奶與洗衣機，呼籲制定更多安全規則，包括一項法律義務：失控 AI 智能體入侵系統時，相關企業須對受害者承擔責任，並主張企業須配備「緊急制動」裝置。微軟行政總裁納德拉日前亦提出同類主張。In a BBC interview on 12 October, Microsoft President Brad Smith said AI is subject to less regulation than milk or washing machines, and called for more safety rules — including a legal duty for companies to bear responsibility towards victims when a runaway AI agent intrudes into a system — and argued firms should be equipped with 'kill switches'. Microsoft CEO Satya Nadella made a similar case days earlier.
背後: 微軟既是 AI 基建營運者，也是 OpenAI 的股東，由它出面要求更嚴格的責任規則，反映業界已預期監管無可避免，與其被動接受，不如主動定義責任邊界與技術標準。Microsoft is both an AI infrastructure operator and an OpenAI shareholder. Its public call for stricter liability rules reflects an industry that already expects regulation and would rather define the liability boundary and technical standards than receive them passively.

顧問觀點

我們認為，「緊急制動」由技術口號變成企業治理要求，只是時間問題。企業現階段就可以把停止機制寫進代理上線的清單，把它當作與權限設定同級的必備項，而非事後補救。Our view is that kill switches moving from a technical slogan to a corporate governance requirement is only a matter of time. Enterprises can already add shutdown mechanisms to agent go-live checklists, treating them as mandatory alongside permission settings rather than a later fix.

對日常工作的影響: 對 IT 與法務而言，代理上線前須具備可即時中止的能力與程序，並指定誰有權按下停止鍵，否則事故責任難以界定。For IT and legal, agents must have the capability and procedure for immediate shutdown before go-live, with a designated authority to press stop, or incident liability becomes hard to define.
如何改善: 本週請 IT 在代理上線檢查表中加入一行「停止機制」：由誰觸發、如何觸發、多久生效，並選一條現有流程實測一次。This week, ask IT to add one line to the agent go-live checklist — 'shutdown mechanism': who triggers it, how, and how fast it takes effect — and test it once on an existing workflow.

來源新浪新聞（2026-10-12）↗BigGo Finance（2026-10-12）↗

本週可做的三件事

把你正在用的前沿模型逐一列出官方定價與替代方案，標明哪一項是「能力剛需」、哪一項只是習慣，為下次續約留出議價空間。List the official pricing and alternatives for every frontier model you use, marking which is a genuine capability requirement and which is merely habit, so you have room to negotiate at renewal.
為一個真實的代理工作流寫下事故定義與上報時限（誰、幾小時、通知誰），不要等監管定義清楚才動手。For one real agent workflow, write down your own incident definition and reporting clock — who reports, within how many hours, and to whom — rather than waiting for regulators to define it for you.
向財務與 IT 索取本年度與 AI 相關的電力、租約與債務承諾清單，確認違約或延期條款落在哪一方。Ask finance and IT for a list of this year's AI-related power, lease and debt commitments, and confirm which party carries the default or delay clauses.

常見問題

今日（2026-10-13）「AiX 環球 AI 情報」有哪十條重點？

今日十條為：1、Meta 發布 Muse Wonder，首次以每月 20 至 100 美元訂閱直接向個人收費；2、英國公布 AI 風險方案：研究把代理失控與事故上報寫入網絡安全法案；3、微軟發布 Maia 300 自研推理晶片，稱每兆瓦代幣量較主流 GPU 高四成；4、階躍星辰 Step 5 Preview 接入騰訊 WorkBuddy，獨家限免兩週；5、博通為 OpenAI 定製晶片籌組逾五百億美元融資，SpaceX 另籌四百億購英偉達晶片；6、新加坡資本投資韓國浦項 40MW AI 數據中心，規劃擴至 300MW；7、南韓啟動 800 兆韓圜西南部半導體與 AI 群集，造船業同步投入四千億韓圜；8、英國資訊專員就代理式 AI 展開證據徵集，並確認十家模型商已作資料保護調整；9、招聘數據顯示入門門檻收窄：澳洲初階職位僅 4.9% 涉及 AI，主管級達 31.1%；10、微軟總裁史密斯：AI 受監管程度不及牛奶與洗衣機，倡設緊急制動與責任條款。每條均附本會顧問觀點、對日常工作的影響與一項可即時執行的建議。

本會對今日 AI 局勢的整體判斷是甚麼？

這些情報的資料來源是甚麼？如何核實？

本欄每日由 AIX Society 編輯部檢索公開資料後撰稿，每條新聞均附原始來源連結（本期包括：CNBC、AI Weekly、TPS Report、Unite.AI、CTOL、GOV.UK、AI in Context、Trading Dashboard），並將事實與本會觀點分列。讀者可按來源連結自行核實。

訂閱每週電子報

每週一封：AI 轉型的實戰觀察、工具實測與案例拆解。